The 2019 filing described one loop: label generated samples, train a personal classifier on their latent vectors, then run it forward to score and backward to generate. Since then the pieces of that loop have been reinvented independently, again and again — none of the work below cites the original, and that's the point. Independent convergence is the strongest evidence an idea is real. Here is the record, dated and sourced, followed by what still doesn't exist.
A thumbs-up/down rating game feeding a latent style space — billions of ratings, millions per day — used to personalize clothing recommendations at industrial scale.
Same: swipe-labeling into a latent preference space, contemporaneous with the SkinDeep.ai work. Different: ranks a real inventory; doesn't generate, and doesn't invert.
Thirty participants viewed GAN-generated faces while EEG recorded their responses. A classifier trained on those brain reactions steered latent-space optimization to generate faces each person would find attractive — validated at roughly 80% in double-blind evaluation. Published in IEEE Transactions on Affective Computing.
Same: the exact mechanism, end to end — generated faces, latent classifier, optimization, decode. The 2019 filing even names brain interfaces as a future rating device. Different: a lab study, not a product; EEG instead of swipes.
Onboarding asks users to rate faces; a model learns their visual type and surfaces candidates predicted to be mutually attracted. Over a million users.
Same: rate-to-train attraction models used for two-way matching — the filing's dating application. Different: operates on real photos in a centralized service; no generative latent space, no privacy-by-encoding.
Users rank image pairs (about 200 to start); Midjourney builds a personal preference model applied at generation with a strength dial — and profiles are shareable, so others can generate with your taste model.
Same: rate examples → personal model → personalized generation, in the mainstream. Shareable models echo the filing's group/other-person classifiers. Different: steers the generation process rather than solving for your optimum; needs hundreds of ratings, processed server-side, not milliseconds on device.
Infers a user's liked and disliked visual attributes from a handful of commented ratings, then conditions a text-to-image model on that profile. Its user studies report the personalized outputs winning overwhelmingly.
Same: per-user preference profiles driving generation, academically formalized. Different: preferences extracted through a language model into conditioning space, not a classifier over the generator's latent space.
Per-user reward models for generative AI now have benchmarks, workshops, and a naming convention: PersonalLLM (ICLR 2025), personalized reward modeling for text-to-image, per-user preference benchmarks, and studies of how personalized reward models should be selected and aligned.
Same: the classifier half of the 2019 filing — small preference models per person, used to steer big generative models. Different: mostly aggregate-then-adapt architectures; inversion to the optimum is still rare.
Feeds of AI-generated video tuned by engagement: OpenAI's Sora app launched September 2025 (and was discontinued in March 2026); Meta's Vibes feed continues, with Meta reporting strong retention and weighing a standalone app. Zero prompting required — you swipe, the feed adapts.
Same: the filing's zero-prompt content feed — generated media selected by learned preference. Different: they rank a pool of generations with engagement signals; they don't yet train an explicit per-user model and steer generation with it.
As of this writing, no mainstream product ships:
Why the gap persists: aggregate preference tuning fits how the industry already ships one big model to everyone, while per-user models are a product-architecture change — storage, cold-start, and interface questions more than research questions. Those are exactly the questions the 2019 system answered at small scale, and the whitepaper and code answer in the open today.